Harishankar Vasudevanallur Subramanian, a doctoral candidate in engineering management, will defend their dissertation titled “Supporting Human-AI Interaction from User Expectations to Mental Models.” Their advisor, Dr. Casey Canfield, is an associate professor in the engineering management and systems engineering department. The dissertation abstract is provided below.

Explainable AI (XAI) aims to unravel the "black-box" nature of AI systems and provide insights into the inner workings that lead to a prediction. However, the best XAI communication varies depending on the individual, task, and broader context. It is challenging to anticipate the best XAI for a particular use case. One strategy is for users to build an appropriate mental model of AI with both prediction and system level XAI. To date, little research has focused on quantitatively measuring users interacting with system-level XAI. This dissertation has three primary contributions. The first contribution is a scoping review paper focused on developing AI for kidney transplant placement identifying the need for user-driven customization capabilities. This work highlights the need for stakeholders’ input when designing an AI system and how much information they desire to review for each case, depending on user expertise, task complexity, and AI literacy. The second contribution focuses on the impact of prediction-level information by measuring user task performance and confidence when they receive multiple AI recommendations along with AI’s uncertainty information. The third contribution is a series of experiments to measure discernment (i.e., people’s ability to evaluate the capability of an AI) to understand a user’s mental model of AI. These studies support efforts to build theory and develop an interface for an AI system that can be integrated into the kidney transplant placement process to reduce the kidney non-utilization rate.